[RISCV] Prevent tryToFoldBNEOnCmpXchgResult from deleting AND if has t others users.

[craig.topper]

Summary

This disables the transform if the branch does not have the kill
flag set for the AND we want to delete.

Ideally we'd be able to share the AND with the AND we create in
the expansion, but that's a more complex transform. So this starts
with the simple approach to fix miscompile.

This should be backported to LLVM 17.

Fixes PR65025.ll

[asb]

LGTM, thanks.

llvm/lib/Target/RISCV/RISCVExpandAtomicPseudoInsts.cpp

bool tryToFoldBNEOnCmpXchgResult(MachineBasicBlock &MBB,
   // Match BNE DestReg, MaskReg.
   if (MBBI == E || MBBI->getOpcode() != RISCV::BNE)
     return false;
   Register BNEOp0 = MBBI->getOperand(0).getReg();
   Register BNEOp1 = MBBI->getOperand(1).getReg();
   if (!(BNEOp0 == DestReg && BNEOp1 == CmpValReg) &&
       !(BNEOp0 == CmpValReg && BNEOp1 == DestReg))
     return false;
+
+  // Make sure the branch is the only user of the AND.
+  if (MaskReg.isValid()) {
+    if (BNEOp0 == DestReg && !MBBI->getOperand(0).isKill())
+      return false;
+    if (BNEOp1 == DestReg && !MBBI->getOperand(1).isKill())
+      return false;
+  }
+
   ToErase.push_back(&*MBBI);
   LoopHeadBNETarget = MBBI->getOperand(2).getMBB();
   MBBI = skipDebugInstructionsForward(std::next(MBBI), E);
   if (MBBI != E)
     return false;
 
   MBB.removeSuccessor(LoopHeadBNETarget);
   for (auto *MI : ToErase)

llvm/test/CodeGen/RISCV/pr65025.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
+; RUN: llc < %s -mtriple=riscv64 -mattr=+a | FileCheck %s
+
+define ptr @cmpxchg_masked_and_branch1(ptr %ptr, i8 signext %cmp, i8 signext %val) nounwind {
+; CHECK-LABEL: cmpxchg_masked_and_branch1:
+; CHECK:       # %bb.0: # %do_cmpxchg
+; CHECK-NEXT:    andi a3, a0, -4
+; CHECK-NEXT:    slli a4, a0, 3
+; CHECK-NEXT:    li a5, 255
+; CHECK-NEXT:    sllw a5, a5, a4
+; CHECK-NEXT:    andi a1, a1, 255
+; CHECK-NEXT:    sllw a1, a1, a4
+; CHECK-NEXT:    andi a2, a2, 255
+; CHECK-NEXT:    sllw a2, a2, a4
+; CHECK-NEXT:  .LBB0_3: # %do_cmpxchg
+; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    lr.w.aqrl a4, (a3)
+; CHECK-NEXT:    and a6, a4, a5
+; CHECK-NEXT:    bne a6, a1, .LBB0_5
+; CHECK-NEXT:  # %bb.4: # %do_cmpxchg
+; CHECK-NEXT:    # in Loop: Header=BB0_3 Depth=1
+; CHECK-NEXT:    xor a6, a4, a2
+; CHECK-NEXT:    and a6, a6, a5
+; CHECK-NEXT:    xor a6, a4, a6
+; CHECK-NEXT:    sc.w.rl a6, a6, (a3)
+; CHECK-NEXT:    bnez a6, .LBB0_3
+; CHECK-NEXT:  .LBB0_5: # %do_cmpxchg
+; CHECK-NEXT:    and a2, a4, a5
+; CHECK-NEXT:    bne a1, a2, .LBB0_2
+; CHECK-NEXT:  # %bb.1: # %returnptr
+; CHECK-NEXT:    xor a1, a1, a2
+; CHECK-NEXT:    snez a1, a1
+; CHECK-NEXT:    addi a1, a1, -1
+; CHECK-NEXT:    and a0, a1, a0
+; CHECK-NEXT:    ret
+; CHECK-NEXT:  .LBB0_2: # %exit
+; CHECK-NEXT:    li a0, 0
+; CHECK-NEXT:    ret
+do_cmpxchg:
+  %0 = cmpxchg ptr %ptr, i8 %cmp, i8 %val seq_cst seq_cst
+  %1 = extractvalue { i8, i1 } %0, 1
+  %2 = select i1 %1, ptr %ptr, ptr null
+  br i1 %1, label %returnptr, label %exit
+returnptr:
+  ret ptr %2
+exit:
+  ret ptr null
+}